CatBoost官方基准测试示例是否使用测试数据作为有效数据？

如何解决CatBoost官方基准测试示例是否使用测试数据作为有效数据？

在检查CatBoost Official Benchmark examples时，我注意到代码将测试数据用作最终验证的“ eval_set”。这是正确的方法吗？我的印象是“测试数据”保留用于最终测试。

来自here的代码。专注于 ITEM 1

def run_experiment(Experiment,title):
    experiment = Experiment(learning_task,max_hyperopt_evals=max_hyperopt_evals,n_estimators=n_estimators)
    cv_pairs,(dtrain,dtest) = experiment.split_and_preprocess(X_train.copy(),y_train,X_test.copy(),y_test,cat_cols,n_splits=5)
 
    default_cv_result = experiment.run_cv(cv_pairs)
    experiment.print_result(default_cv_result,'Default {} result on cv'.format(title))

    print('\n~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\n')

    default_test_losses = []
    for seed in range(5):
        test_result = experiment.run_test(dtrain,dtest,X_test,params=default_cv_result['params'],n_estimators=default_cv_result['best_n_estimators'],seed=seed)
        default_test_losses.append(test_result['loss'])
        print 'For seed=%d Test\'s %s : %.5f' % (seed,experiment.metric,default_test_losses[-1])
    print '\nTest\'s %s mean: %.5f,Test\'s %s std: %.5f' % (experiment.metric,np.mean(default_test_losses),np.std(default_test_losses))

    print('\n~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\n')
    print('Hyperopt iterations:\n\n')

    tuned_cv_result = experiment.optimize_params(cv_pairs)

    print('\n~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\n')

    experiment.print_result(tuned_cv_result,'Tuned {} result on cv'.format(title))

    print('\n~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\n')

    tuned_test_losses = []
    for seed in range(5):
        ####################################################
        # ITEM 1
        ########################################################
        test_result = experiment.run_test(dtrain,params=tuned_cv_result['params'],n_estimators=tuned_cv_result['best_n_estimators'],seed=seed)
        tuned_test_losses.append(test_result['loss'])
        print 'For seed=%d Test\'s %s : %.5f' % (seed,tuned_test_losses[-1])
    print '\nTest\'s %s mean: %.5f,np.mean(tuned_test_losses),np.std(tuned_test_losses))

    return np.mean(default_test_losses),np.mean(tuned_test_losses)

'experiment.py'中的run_test函数：着重于项目＃2

    def run_test(self,dtrain,X_test=None,params=None,n_estimators=None,custom_metric=None,seed=0):
        params = params or self.best_params or self.default_params
        n_estimators = n_estimators or self.best_n_estimators or self.n_estimators
        params = self.preprocess_params(params)
        start_time = time.time()
        ####################################################
        # ITEM 2
        ########################################################
        bst,evals_result = self.fit(params,n_estimators,seed=seed)
        eval_time = time.time() - start_time
        preds = self.predict(bst,X_test)

        result = {'loss': evals_result[-1],'bst': bst,'n_estimators': n_estimators,'eval_time': eval_time,'status': STATUS_OK,'params': params.copy(),'preds': preds}

        if custom_metric is not None:
            if type(custom_metric) is not dict:
                raise TypeError("custom_metric argument should be dict")
            pred = self.predict(bst,X_test)
            for title,func in custom_metric.iteritems():
                score = func(dtest.get_label(),pred,sample_weight=None) # TODO weights
                result[title] = score

        return result

'catboost_experiment.py'中catboost_experiment.py中的fit函数。专注于项目3

    def fit(self,params,seed=0):
        params.update({"iterations": n_estimators})
        params.update({"random_seed": seed})
        bst = CatBoost(params)
        ####################################################
        # ITEM 3
        ########################################################
        bst.fit(dtrain,eval_set=dtest)
        with open("test_error.tsv","r") as f:
            results = np.array(map(lambda x: float(x.strip().split()[-1]),f.readlines()[1:]))
        
        return bst,results

代码使用dtest集作为validation集。这是正确的方法吗？

CatBoost fit使用eval_set作为“过拟合检测器”，“最佳迭代选择”和“监控指标”更改的验证数据集。

我很困惑，如果我缺少一些东西，请纠正我。